Meta open-sources a distillation of its closed flagship: Muse Glimmer runs agents on one 24 GB card

Meta published Muse Glimmer on 10 August 2026 — a dense model of about 29.6 billion parameters that the company describes as trained on the outputs of its closed Muse Spark flagship by logit distillation. The weights are on Hugging Face under a plain Apache 2.0 licence, without the usage thresholds that Meta's earlier community licences carried. The stated design goal is "always-on local agent workflows": a model that stays resident on a personal machine, calls tools, writes and debugs code, reads screenshots and works through multi-step tasks without a network round trip. Vision comes from a frozen ViT-G/14 encoder of roughly 1.8 billion parameters, and the context window is 131,072 tokens — well short of the million offered by the teacher model, which is the trade for the smaller size. The hardware numbers are the substance of the announcement. Full precision needs 64 GB, but Meta ships two four-bit quantisations: 32 GB at a stated 0.2% quality cost, and 24 GB — a single RTX 5090 — at about 1%. Shipped with the model is DFlash, a speculative-decoding drafter head that Meta measures at 233 tokens per second on that card versus 75 without it, and 1.5 to 1.8 times gains on MacBook M4-Max and M5-Max. Integrations for llama.cpp, MLX, ExecuTorch, Ollama and LM Studio were announced alongside. On its own model card Meta reports 76.0 on SWE-Bench Verified, 51.2 on SWE-Bench Pro, 65.9 on OSWorld-Verified, 94.7 on AIME 2026 and 83.5 on GPQA Diamond. All of these are vendor figures; no independent index had published a score at the time of writing. Meta says the open release was assessed under its Advanced AI Scaling Framework and publishes no external audit.